OpenAI put GPT-6.1 Terra on sale October 26 at half the price of the model it replaces. The pitch is narrow and practical: a model that reads invoices, fills in forms and clicks through desktop apps without billing like a flagship.
What OpenAI Shipped With GPT-6.1 Terra
OpenAI published GPT-6.1 Terra on October 26 and framed it as the fast, cheap member of the family rather than the smartest one. The company says it combines detailed visual reasoning, precise computer use and clearer writing, and that it replaces GPT-5.6 Terra, the previous entry in the same line of cheaper, faster releases.
Availability started the same day. Plus, Pro, Enterprise and Edu subscribers get it inside ChatGPT Work and Codex, and developers can call it through the API as gpt-6.1-terra. There's no waitlist and no separate tier to buy.
That detail matters more than it sounds. A cheaper model that lands everywhere at once changes budgets immediately, while a research preview changes nothing for anyone outside a lab.
GPT-6.1 Terra Pricing: $1 In, $5 Out, Cheaper Cache
The headline rate is $1 per million input tokens and $5 per million output tokens on the standard API tier. Cached input falls to $0.05. According to OpenAI, those prices are roughly half what GPT-5.6 Terra and GPT-6.1 Sol charge.
Artificial Analysis added a number OpenAI's page skips: cache writes also halved, to $1.25 per million tokens. If you fire the same long prompt thousands of times a day, the cached read price decides whether the job is viable at all.
GPT-6.1 Terra API | Price per million tokens |
|---|---|
Input | $1.00 |
Cached input | $0.05 |
Cache write | $1.25 |
Output | $5.00 |
Read the output price twice. Output tokens are where agents burn money, because every tool call, every reasoning step and every rewritten file comes back as output. Halving that rate cuts the bill on exactly the workloads that were too expensive to automate last quarter. That's where the savings land.
Where Terra Sits Among Astra, Sol and Luna
OpenAI now sells four rungs. Astra is the strongest, Sol is the balanced one, Terra is the cheap workhorse, and Luna is the budget option. GPT-6.1 Terra isn't a flagship, and the announcement never pretends otherwise.
The company also says the same subscription allowance stretches further on Terra, roughly twice as many tasks as Sol by the estimate an OpenAI staffer posted on X. Treat that as a company claim until your own usage dashboard agrees.
The split is easy to apply in practice. Send hard coding and research work to Sol or Astra. Move the dull, high-volume work down to Terra and stop paying flagship prices for form entry.
OpenAI's Own Results on Computer Use
OpenAI reports its biggest jump on Agents' Last Exam, a benchmark of professional tasks that includes turning architectural drawings into 3D models and logging an animated film shot by shot. At medium reasoning effort, the company says Terra beats GPT-5.6 Terra by 15 percentage points at about a third of the cost per task.
On OSWorld 2.1, which tests multi-step work inside desktop apps, OpenAI says medium effort tops the best score GPT-5.6 Terra reached at any setting. On MMMU-Pro, a visual reasoning test, OpenAI says Terra passes its predecessor at every effort level.
All three are vendor numbers. Nobody outside OpenAI has rerun them on identical tasks, and benchmark versions drift between launches.
The Artificial Analysis Numbers, Which Are Third Party
Artificial Analysis tested the model on its own and reached conclusions that are more mixed than OpenAI's page. At max effort, Terra gains 6 points on the Intelligence Index over GPT-5.6 Terra while cost per task falls from $1.40 to $0.40. At low and medium effort the gain is larger, between 7 and 13 points.
Here's the part the launch post leaves out. Only the low and medium effort settings sit on the Pareto frontier for intelligence against cost. The higher settings land behind it, which means you pay extra for a score another model already reaches cheaper.
The Coding Agent Index v2.0 tells a similar story. Terra scores 48, six points under GPT-6.1 Sol and 17 above GPT-6 Luna. Its xhigh and low settings sit on the cost efficiency frontier, and the middle of the range doesn't.
How GPT-6.1 Terra Cut Hallucinations and Lifted Terminal-Bench 4.0
The most useful change may be the least dramatic. On OpenAI's internal test of error-inducing prompts, the error rate at low effort fell from 53.6% to 16.7%, and at max effort to 4.3%. That's a drop of nearly 37 points at the low end.
Artificial Analysis measured something adjacent. In AA-Omniscience, Terra guesses when it doesn't know the answer 38% of the time, down from 88% for GPT-5.6 Terra, while accuracy rose 2 points. A model that says "I can't read this scan" is worth more in an accounting pipeline than one that invents a number.
On Terminal-Bench 4.0, which tests command line problem solving, OpenAI reports 50.3% at max effort against 21.5% published for GPT-5.6 Terra. Artificial Analysis logged a gain of about 20 points, roughly in line with the company's account.
GPT-6.1 Terra Safety: Critical on Cyber
OpenAI says Terra reaches the Critical threshold for cybersecurity under its Preparedness Framework and ships with the protections built for GPT-6 Astra. The system card carries the details and the caveats.
On prompt injection, the company says the model resists instructions designed to pull it off task, and that it more often states its limits instead of guessing. Those results come from internal testing, which is normal on launch day and worth reading as a claim.
A cheap model that clicks around your desktop is also a cheap model that can be talked into clicking the wrong thing. Lower prices make it easier to deploy, and the threat model doesn't get simpler because the invoice is smaller.
What GPT-6.1 Terra at Half Price Actually Changes
The interesting question isn't whether Terra beats Astra. It doesn't come close, and OpenAI says so in the post. What changes at that price?
Boring work becomes affordable. A form that costs a fraction of a cent to process is a form nobody needs to type twice, and a clip that takes four seconds to trim is a clip nobody queues by hand.
Invoice entry, ticket triage, drawing review, footage logging. Terra is aimed at that shelf, and the price is the argument.
Pick it when volume matters and mistakes are recoverable. Keep Sol or Astra for anything where a wrong answer costs more than the tokens. And check which effort setting you're on before you complain about quality, since the cheap end is where this model earns its price.
